Explainable machine learning for predicting tensile properties of aluminum alloys in the laser powder bed fusion process

Abstract

The rapid solidification and unique thermal gradients inherent to the laser powder bed fusion of metals (PBF-LB/M) process limit the suitability of conventional aluminum (Al) alloys, necessitating the optimization of existing alloys or the development of new compositions to achieve the desired tensile properties while ensuring good processability. Experimental exploration of alloy compositions is labor-intensive, costly, and time-consuming. Machine learning (ML) offers a cost-effective, flexible approach to streamline alloy design and accelerate advancements in AM technologies. This study introduces a data-driven predictive framework for predicting tensile properties of Al alloys for PBF-LB/M. To address the limited data on LPBF of Al alloys and the restricted range of alloy systems investigated, data of conventional Al alloys (including cast and wrought alloys) and laser-directed energy deposition (DED-LB/M) built Al alloys were also included, alongside PBF-LB/M data. The dataset incorporates a comprehensive pool of features such as alloy composition, processing parameters, grain size, and elemental properties. The Pearson correlation coefficient (PCC) with feature importance-based feature selection was implemented to balance model complexity and accuracy via reducing the dimensionality and overfitting. The resulting ML framework demonstrates excellent predictive accuracy and generalizability, successfully extending its applicability to unseen alloy systems. This framework offers a reliable tool for optimizing Al alloy designs, significantly reducing reliance on costly experimental trials. The inclusion of Explainable AI provided detailed interpretability, elucidating the influence of individual features on model predictions, ensuring the predictions were scientifically grounded.